An explainable machine learning framework utilizing ultrasound radiomics for the preoperative differentiation between...

To develop an interpretable machine learning model utilizing ultrasound radiomics for distinguishing between granulomatous lobular mastitis and breast cancer.

This retrospective study encompassed 237 patients who underwent preoperative breast ultrasound examinations and received pathological diagnoses of either granulomatous lobular mastitis or breast cancer at Quzhou People's Hospital between April 2013 and April 2023. Radiomic features were extracted from the ultrasound images, and feature selection was conducted using intra-class correlation coefficients, Pearson correlation coefficients, and the least absolute shrinkage and selection operator regression. Machine learning models based on radiomics were constructed using Extremely Randomized Trees, Light Gradient Boosting Machine, and Random Forest. Additionally, a combined model was developed by integrating independent clinical predictors with the radiomics signature.The model's performance was assessed using the area under the receiver operating characteristic curve, accuracy, sensitivity, and specificity. To evaluate clinical utility, decision curve analysis was employed, while Shapley additive explanation was utilized to interpret model explainability.

A total of 1,161 radiomic features were extracted from each ultrasound image. Following Pearson correlation filtering, 135 features were retained, and 15 features were selected using the least absolute shrinkage and selection operator regression for model construction. The combined model, which integrated clinical factors with the radiomics signature, exhibited superior performance, achieving an AUC of 0.935 (95% CI: 0.902-0.969) in the training cohort and 0.833 (95% CI: 0.710-0.950) in the validation cohort. DCA indicated favorable clinical applicability. The Shapley additive explanation analysis shows that the imaging biomarker features lbp_3D_k_glszm_SmallAreaLowGrayLevelEmphasis, gradient_glcm_Imc2, gradient_glszm_ZoneEntropy, original_shape_Elongation, squareroot_glrlm_RunEntropy, and wavelet_HLH_glszm_LowGrayLevelZoneEmphasis have a strong correlation with the prediction of granulomatous lobular mastitis.

The combined model, which incorporates ultrasound radiomics and clinical factors, exhibited significant efficacy in preoperatively differentiating granulomatous lobular mastitis from breast cancer. This non-invasive and interpretable methodology shows potential for enhancing diagnostic precision and guiding clinical decision-making.